RON: Reverse Connection with Objectness Prior Networks for Object Detection

Transcription

1 RON: Reverse Connection with Objectness Prior Networks for Object Detection Tao Kong 1, Fuchun Sun 1, Anbang Yao 2, Huaping Liu 1, Ming Lu 3, Yurong Chen 2 1 Department of CST, Tsinghua University, 2 Intel Labs China 3 Department of EE, Tsinghua University

2

3 mean Average Precision (map) The progress of object detection PASCAL VOC % 80% 60% 88.4 Region Based. R-CNN, Fast(er) R-CNN, R-FCN ION, HyperNet, OHEM, MR-CNN, etc. 40% 20% Region Free. YOLO, SSD etc. 0% year

4 Two object detection architectures SSD a) featurized image pyramid b) Single-shot detector with on CNN c) Multiple anchors at one level feature 1.High-speed 2.No repeated computations 3.Not easy to train 4.Strugle with small instances Faster R-CNN a) Region proposal network b) Region-wise object detection sub-network 1.High-accuracy, easy to train 2.Easy to follow 3.Resource/time consuming 4.Repeated computation with region-wise computing

5 So, what is the lesson? Feature pyramid works better in locating all scales of objects from: SSD, HyperNet, ION, MR-CNN Using region proposal network to reduce searching space from: R-FCN, Faster R-CNN, Fast R-CNN RON A fully CNN pipeline with no repeated computation can achieve high detection performance. from: SSD, R-FCN

6 So, what is the lesson? Feature pyramid works better in locating all scales of objects from: SSD, HyperNet, ION, MR-CNN Using region proposal network to reduce searching space from: R-FCN, Faster R-CNN, Fast R-CNN RON A fully CNN pipeline with no repeated computation can achieve high detection performance. from: SSD, R-FCN

7 RON: Reverse Connection with Objectness Prior Networks Objectness Prior Reverse Connection

8 What is reverse connection and why? detect detect detect Single feature map ((Fast(er) R-CNN) detect Exploring multiple feature directly (SSD) detect Multiple feature concatenation (ION, HyperNet) detect detect detect Multiple feature with reverse connection

9 What is reverse connection and why? a) Simple design b) The semantic information of former layers can be significantly enriched c) Keep the spatial sizes detect d) Easy to do multiple level detection detect detect A reverse connection block

10 So, what is the lesson? Feature pyramid works better in locating all scales of objects from: SSD, HyperNet, ION, MR-CNN Using region proposal network to reduce searching space from: R-FCN, Faster R-CNN, Fast R-CNN RON A fully CNN pipeline with no repeated computation can achieve high detection performance. from: SSD, R-FCN

11 From region propsoal boxes to region proposal maps Faster R-CNN: Region proposal network (RPN) is fully convolutinal, but detection subnetwork is with repeated computation (ROI-Pooling). Why? The bbox regression in RPN changes the spatial locations of all boxes, which breaks the anchor's reletionship with its corresponding kernel.

12 From region propsoal boxes to region proposal maps Objectness pior: Share anchors between RPN and detector, make it posible to detect objects with fully ConvNet. No repeated computations, much faster The total network is fully convolutional There are one map for each type of anchors different from these mask-based methods.

13 Region proposal maps

14 So, what is the lesson? Feature pyramid works better in locating all scales of objects from: SSD, HyperNet, ION, MR-CNN Using region proposal network to reduce searching space from: R-FCN, Faster R-CNN, Fast R-CNN RON A fully CNN pipeline with no repeated computation can achieve high detection performance. from: SSD, R-FCN

15 RON bounding box generation 4*4 feature map Convolutinal output in this position a) sub-network for objectness b) sub-network for detection with a) c) sub-network for bounding box regression 8*8 feature map

16 RON object detector Object detection and bounding box regression modules. Top: bounding box regression; Bottom: object classification

17 RON optimization objectness prior bbox location detection optimize the network

18 Main results +2.5% +2.3%

19 Main results +3.7%

20 Main results

21 Main results +20%

22 Tao Kong, Fuchun Sun, Anbang Yao, Huaping Liu, Ming Lu, Yurong Chen. RON: Reverse Connection with Objectness Prior Networks for Object Detection, In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Paper: Check out the code/models

23 Thanks